A multi-modal machine learning classifier for the differential diagnosis of inflammatory neuropathies and demyelinating diseases: A diagnostic accuracy study

Background and Aims The differential diagnosis of inflammatory neuropathies—including Acute Inflammatory Demyelinating Polyneuropathy (AIDP), Chronic Inflammatory Demyelinating Polyneuropathy (CIDP), Multiple Sclerosis (MS), Pseudotumor Cerebri (PTS), and Neuromyelitis Optica Spectrum Disorder (NMOSD)—remains a clinical challenge. Current diagnostic paradigms rely on integrating clinical presentation, cerebrospinal fluid (CSF) analysis, electrophysiology, and neuroimaging, which can lead to diagnostic delays and misclassification. We aimed to develop and validate a multimodal machine-learning classifier to accurately distinguish between these conditions using routinely available clinical data. Methods In this diagnostic accuracy study, we analyzed data from 438 patients (54.3% female; mean age 51.2 ± 16.8 years) with confirmed diagnoses of AIDP (n=104), CIDP (n=89), MS (n=112), PTS (n=78), and NMOSD (n=55). A multimodal ensemble classifier integrating gradient-boosted trees and a neural network was developed to process 47 features across five domains: clinical demographics; CSF biomarkers (IgG index, QAlb, NfL, GFAP, protein); serum biomarkers (homocysteine, CRP); electrophysiology (nerve conduction studies); and neuroimaging findings. Model performance was evaluated using five-fold cross-validation and an independent external validation cohort (n=88) from a separate tertiary center. The primary outcome was diagnostic accuracy, with secondary outcomes including area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and calibration. Results The multi-modal classifier achieved a macro-average AUC of 0.94 (95% CI 0.91–0.97) in cross-validation and 0.91 (95% CI 0.86–0.95) in external validation. Overall diagnostic accuracy was 87.2% (95% CI 83.1–91.3). For individual diagnoses, sensitivity ranged from 0.83 (PTS) to 0.94 (MS), and specificity from 0.89 (AIDP) to 0.97 (NMOSD). CSF neurofilament light chain (NfL), IgG index, and tibial nerve conduction velocity emerged as the most influential features. The classifier significantly outperformed clinician diagnosis alone (accuracy 71.4% vs. 87.2%; p<0.001) and demonstrated excellent calibration (Brier score 0.082). Conclusion A multi-modal machine learning classifier accurately distinguishes between common inflammatory neuropathies and demyelinating diseases. This approach has the potential to reduce diagnostic uncertainty, accelerate treatment initiation, and serve as a clinical decision support tool, particularly in settings with limited access to subspecialist expertise. Prospective validation is warranted before clinical implementation.